Efficiency improvement of Thai-English hierarchical phrase-based translation system using adaptive translation table with bilingual dictionary

การเพิ่มประสิทธิภาพเครื่องแปลภาษาไทยเป็นภาษาอังกฤษแบบลำดับขั้นวลีโดยใช้การปรับตารางการแปลภาษาด้วยพจนานุกรมสองภาษา

Name: Prasert Luekhong

LCSH: Machine translating

Classification :.DDC: 418.020285

Abstract: Machine translation is the software to translate text or speech from one language to another language. A decade ago, the trend was statistical approach that the efficient of that depends on the models of the training processes. The models of statistical machine translation (SMT) were the language model and translation model. However, the results of the incorrect translation have been occurred from previous works mostly pointing to the efficiency of word alignment in the training processes. This issue became worse for the explicitly disputative pattern of language pairs such as English-Thai translation, which belongs to a different language family. Moreover, the polysemous words and words shorten in forms that were rich in Thai natural language often leaded to be wrong corelate statistic. The assumption of this thesis was the prior knowledge of linguistic can be improved the accuracy of the statistical word alignment process. However, what were the suitability SMT models between phrase-based translation (PBT), and hierarchical phrasebased translation (HPBT) for Thai-based SMT. Thus, this thesis aims to comparative study the HPBT and PBT in Thai-based translation, to propose the improvement of training process with bilingual dictionary for Thai-English statistical machine translation. Firstly, this thesis studied on applying 3, 4, 5 and 6-gram PBT model and 3, 4, 5 and 6- gram HPBT model to translate Thai-English translation using 10 fold cross-validation viii experiments. The results showed that HPBT higher than PBT for each language model. When n-gram was 3 got higher than PBT 1.65 BLEU Score, 4 gram was 1.16 BLEU Score, and 5 gram was 1.15 BLEU Score in Thai-to-English translation. Furthermore, this thesis studied on applying HPBT and PBT model to translate Thai-Chinese translation. The results showed that HPBT higher than PBT when n-gram was 3 gram in Thai-to-Chinese and Chinese-to-Thai. The above preliminary experiment ensure that the suitable statistical translation approach for Thai-based SMT was HPBT. However, this thesis observations showed that the majority problem of poor results of Thai-based SMT were the misaligned and unknown words according to this model only used the cooccurrence word alignment. So, this thesis aim to improve the training process with bilingual dictionary for Thai-based SMT. Next, this thesis proposed the alignment framework employs the benefit of both cooccurrence and prior knowledge from a dictionary. The system has been designed to work separately as two steps. The first step was the dictionary-based step to guarantee the accurate word-aligning and the second step was the co-occurrence-based step to handle the unknown word. In the dictionary-based step, the similarity score was exploited to handle effectively partial unknown words. By testing the proposed system against a renowned existing parallel word alignment, GIZA, we applied an alignment model from both systems to Moses for proving its usefulness in a practical machine translation usage and exploited a BLEU score as a measurement. The Thai-to-English translation results showed that the best setting of the proposed method could overrun the widely used word alignment, GIZA IBM model-4, by 1.12 BLEU Score. Finally, this thesis proposed the adaptive translation table with Bilingual Dictionary model that was the lexical probabilistic tuning methodology for translation table in SMT. This methodology used the exactly matching words in bilingual sentence from the dictionary-based alignment to construct the conditional probability of source and target word pair. This process produced a translation table based on the bilingual dictionary. The output table used to weight the lexical probabilistic in baseline translation table. The case study in this work focused on Thai to English translation. From the results, our adaptive translation showed the potential translation result over the baseline HPBT by has better BLEU point for 2.39 BLEU Score. Moreover, the well-known translation software ix Google and Bing got lower results than baseline 4.01 and 6.81 BLEU Score respectively. Due to their using different data sets. Forevermore, our adaptive translation methodology was overcome the Google 6.4 BLEU Score and defeat the Bing 9.2 BLEU Score.

Chiang Mai University Library

Address: Chiang Mai

Email: cmulibref@cmu.ac.th

Name: Thanaruk Theeramunkong

Role: Chairman

Name: Thepchai Supriithi

Role: Member

Name: Jeerayut Chaijaruwanich

Role: Member

Name: Rattasit Sukhahuta

Role: Member

Name: Samerkae Somhom

Role: Member

Created: 2016

Modified: 2020-01-10

Issued: 2020-01-10

วิทยานิพนธ์/Thesis

application/pdf

Th 418.020285 P911E

eng

DegreeName: Doctor of Philosophy

Level: Doctoral Degree

Descipline: Computer Science

Grantor: Chiang Mai University

©copyrights Chiang Mai University

ลำดับที่.	ชื่อแฟ้มข้อมูล	ขนาดแฟ้มข้อมูล	จำนวนเข้าถึง	วัน-เวลาเข้าถึงล่าสุด
1	comp80516prlu_full.pdf	4.55 MB	12	2026-05-26 18:16:01

ใช้เวลา

0.036683 วินาที